drivers/infiniband/hw/ipath/ipath_iba6120.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33 /*
  34  * This file contains all of the code that is specific to the
  35  * InfiniPath PCIe chip.
  36  */
  37
  38 #include <linux/interrupt.h>
  39 #include <linux/pci.h>
  40 #include <linux/delay.h>
  41
  42
  43 #include "ipath_kernel.h"
  44 #include "ipath_registers.h"
  45
  46 /*
  47  * This file contains all the chip-specific register information and
  48  * access functions for the QLogic InfiniPath PCI-Express chip.
  49  *
  50  * This lists the InfiniPath registers, in the actual chip layout.
  51  * This structure should never be directly accessed.
  52  */
  53 struct _infinipath_do_not_use_kernel_regs {
  54         unsigned long long Revision;
  55         unsigned long long Control;
  56         unsigned long long PageAlign;
  57         unsigned long long PortCnt;
  58         unsigned long long DebugPortSelect;
  59         unsigned long long Reserved0;
  60         unsigned long long SendRegBase;
  61         unsigned long long UserRegBase;
  62         unsigned long long CounterRegBase;
  63         unsigned long long Scratch;
  64         unsigned long long Reserved1;
  65         unsigned long long Reserved2;
  66         unsigned long long IntBlocked;
  67         unsigned long long IntMask;
  68         unsigned long long IntStatus;
  69         unsigned long long IntClear;
  70         unsigned long long ErrorMask;
  71         unsigned long long ErrorStatus;
  72         unsigned long long ErrorClear;
  73         unsigned long long HwErrMask;
  74         unsigned long long HwErrStatus;
  75         unsigned long long HwErrClear;
  76         unsigned long long HwDiagCtrl;
  77         unsigned long long MDIO;
  78         unsigned long long IBCStatus;
  79         unsigned long long IBCCtrl;
  80         unsigned long long ExtStatus;
  81         unsigned long long ExtCtrl;
  82         unsigned long long GPIOOut;
  83         unsigned long long GPIOMask;
  84         unsigned long long GPIOStatus;
  85         unsigned long long GPIOClear;
  86         unsigned long long RcvCtrl;
  87         unsigned long long RcvBTHQP;
  88         unsigned long long RcvHdrSize;
  89         unsigned long long RcvHdrCnt;
  90         unsigned long long RcvHdrEntSize;
  91         unsigned long long RcvTIDBase;
  92         unsigned long long RcvTIDCnt;
  93         unsigned long long RcvEgrBase;
  94         unsigned long long RcvEgrCnt;
  95         unsigned long long RcvBufBase;
  96         unsigned long long RcvBufSize;
  97         unsigned long long RxIntMemBase;
  98         unsigned long long RxIntMemSize;
  99         unsigned long long RcvPartitionKey;
 100         unsigned long long Reserved3;
 101         unsigned long long RcvPktLEDCnt;
 102         unsigned long long Reserved4[8];
 103         unsigned long long SendCtrl;
 104         unsigned long long SendPIOBufBase;
 105         unsigned long long SendPIOSize;
 106         unsigned long long SendPIOBufCnt;
 107         unsigned long long SendPIOAvailAddr;
 108         unsigned long long TxIntMemBase;
 109         unsigned long long TxIntMemSize;
 110         unsigned long long Reserved5;
 111         unsigned long long PCIeRBufTestReg0;
 112         unsigned long long PCIeRBufTestReg1;
 113         unsigned long long Reserved51[6];
 114         unsigned long long SendBufferError;
 115         unsigned long long SendBufferErrorCONT1;
 116         unsigned long long Reserved6SBE[6];
 117         unsigned long long RcvHdrAddr0;
 118         unsigned long long RcvHdrAddr1;
 119         unsigned long long RcvHdrAddr2;
 120         unsigned long long RcvHdrAddr3;
 121         unsigned long long RcvHdrAddr4;
 122         unsigned long long Reserved7RHA[11];
 123         unsigned long long RcvHdrTailAddr0;
 124         unsigned long long RcvHdrTailAddr1;
 125         unsigned long long RcvHdrTailAddr2;
 126         unsigned long long RcvHdrTailAddr3;
 127         unsigned long long RcvHdrTailAddr4;
 128         unsigned long long Reserved8RHTA[11];
 129         unsigned long long Reserved9SW[8];
 130         unsigned long long SerdesConfig0;
 131         unsigned long long SerdesConfig1;
 132         unsigned long long SerdesStatus;
 133         unsigned long long XGXSConfig;
 134         unsigned long long IBPLLCfg;
 135         unsigned long long Reserved10SW2[3];
 136         unsigned long long PCIEQ0SerdesConfig0;
 137         unsigned long long PCIEQ0SerdesConfig1;
 138         unsigned long long PCIEQ0SerdesStatus;
 139         unsigned long long Reserved11;
 140         unsigned long long PCIEQ1SerdesConfig0;
 141         unsigned long long PCIEQ1SerdesConfig1;
 142         unsigned long long PCIEQ1SerdesStatus;
 143         unsigned long long Reserved12;
 144 };
 145
 146 #define IPATH_KREG_OFFSET(field) (offsetof(struct \
 147     _infinipath_do_not_use_kernel_regs, field) / sizeof(u64))
 148 #define IPATH_CREG_OFFSET(field) (offsetof( \
 149     struct infinipath_counters, field) / sizeof(u64))
 150
 151 static const struct ipath_kregs ipath_pe_kregs = {
 152         .kr_control = IPATH_KREG_OFFSET(Control),
 153         .kr_counterregbase = IPATH_KREG_OFFSET(CounterRegBase),
 154         .kr_debugportselect = IPATH_KREG_OFFSET(DebugPortSelect),
 155         .kr_errorclear = IPATH_KREG_OFFSET(ErrorClear),
 156         .kr_errormask = IPATH_KREG_OFFSET(ErrorMask),
 157         .kr_errorstatus = IPATH_KREG_OFFSET(ErrorStatus),
 158         .kr_extctrl = IPATH_KREG_OFFSET(ExtCtrl),
 159         .kr_extstatus = IPATH_KREG_OFFSET(ExtStatus),
 160         .kr_gpio_clear = IPATH_KREG_OFFSET(GPIOClear),
 161         .kr_gpio_mask = IPATH_KREG_OFFSET(GPIOMask),
 162         .kr_gpio_out = IPATH_KREG_OFFSET(GPIOOut),
 163         .kr_gpio_status = IPATH_KREG_OFFSET(GPIOStatus),
 164         .kr_hwdiagctrl = IPATH_KREG_OFFSET(HwDiagCtrl),
 165         .kr_hwerrclear = IPATH_KREG_OFFSET(HwErrClear),
 166         .kr_hwerrmask = IPATH_KREG_OFFSET(HwErrMask),
 167         .kr_hwerrstatus = IPATH_KREG_OFFSET(HwErrStatus),
 168         .kr_ibcctrl = IPATH_KREG_OFFSET(IBCCtrl),
 169         .kr_ibcstatus = IPATH_KREG_OFFSET(IBCStatus),
 170         .kr_intblocked = IPATH_KREG_OFFSET(IntBlocked),
 171         .kr_intclear = IPATH_KREG_OFFSET(IntClear),
 172         .kr_intmask = IPATH_KREG_OFFSET(IntMask),
 173         .kr_intstatus = IPATH_KREG_OFFSET(IntStatus),
 174         .kr_mdio = IPATH_KREG_OFFSET(MDIO),
 175         .kr_pagealign = IPATH_KREG_OFFSET(PageAlign),
 176         .kr_partitionkey = IPATH_KREG_OFFSET(RcvPartitionKey),
 177         .kr_portcnt = IPATH_KREG_OFFSET(PortCnt),
 178         .kr_rcvbthqp = IPATH_KREG_OFFSET(RcvBTHQP),
 179         .kr_rcvbufbase = IPATH_KREG_OFFSET(RcvBufBase),
 180         .kr_rcvbufsize = IPATH_KREG_OFFSET(RcvBufSize),
 181         .kr_rcvctrl = IPATH_KREG_OFFSET(RcvCtrl),
 182         .kr_rcvegrbase = IPATH_KREG_OFFSET(RcvEgrBase),
 183         .kr_rcvegrcnt = IPATH_KREG_OFFSET(RcvEgrCnt),
 184         .kr_rcvhdrcnt = IPATH_KREG_OFFSET(RcvHdrCnt),
 185         .kr_rcvhdrentsize = IPATH_KREG_OFFSET(RcvHdrEntSize),
 186         .kr_rcvhdrsize = IPATH_KREG_OFFSET(RcvHdrSize),
 187         .kr_rcvintmembase = IPATH_KREG_OFFSET(RxIntMemBase),
 188         .kr_rcvintmemsize = IPATH_KREG_OFFSET(RxIntMemSize),
 189         .kr_rcvtidbase = IPATH_KREG_OFFSET(RcvTIDBase),
 190         .kr_rcvtidcnt = IPATH_KREG_OFFSET(RcvTIDCnt),
 191         .kr_revision = IPATH_KREG_OFFSET(Revision),
 192         .kr_scratch = IPATH_KREG_OFFSET(Scratch),
 193         .kr_sendbuffererror = IPATH_KREG_OFFSET(SendBufferError),
 194         .kr_sendctrl = IPATH_KREG_OFFSET(SendCtrl),
 195         .kr_sendpioavailaddr = IPATH_KREG_OFFSET(SendPIOAvailAddr),
 196         .kr_sendpiobufbase = IPATH_KREG_OFFSET(SendPIOBufBase),
 197         .kr_sendpiobufcnt = IPATH_KREG_OFFSET(SendPIOBufCnt),
 198         .kr_sendpiosize = IPATH_KREG_OFFSET(SendPIOSize),
 199         .kr_sendregbase = IPATH_KREG_OFFSET(SendRegBase),
 200         .kr_txintmembase = IPATH_KREG_OFFSET(TxIntMemBase),
 201         .kr_txintmemsize = IPATH_KREG_OFFSET(TxIntMemSize),
 202         .kr_userregbase = IPATH_KREG_OFFSET(UserRegBase),
 203         .kr_serdesconfig0 = IPATH_KREG_OFFSET(SerdesConfig0),
 204         .kr_serdesconfig1 = IPATH_KREG_OFFSET(SerdesConfig1),
 205         .kr_serdesstatus = IPATH_KREG_OFFSET(SerdesStatus),
 206         .kr_xgxsconfig = IPATH_KREG_OFFSET(XGXSConfig),
 207         .kr_ibpllcfg = IPATH_KREG_OFFSET(IBPLLCfg),
 208
 209         /*
 210          * These should not be used directly via ipath_read_kreg64(),
 211          * use them with ipath_read_kreg64_port()
 212          */
 213         .kr_rcvhdraddr = IPATH_KREG_OFFSET(RcvHdrAddr0),
 214         .kr_rcvhdrtailaddr = IPATH_KREG_OFFSET(RcvHdrTailAddr0),
 215
 216         /* The rcvpktled register controls one of the debug port signals, so
 217          * a packet activity LED can be connected to it. */
 218         .kr_rcvpktledcnt = IPATH_KREG_OFFSET(RcvPktLEDCnt),
 219         .kr_pcierbuftestreg0 = IPATH_KREG_OFFSET(PCIeRBufTestReg0),
 220         .kr_pcierbuftestreg1 = IPATH_KREG_OFFSET(PCIeRBufTestReg1),
 221         .kr_pcieq0serdesconfig0 = IPATH_KREG_OFFSET(PCIEQ0SerdesConfig0),
 222         .kr_pcieq0serdesconfig1 = IPATH_KREG_OFFSET(PCIEQ0SerdesConfig1),
 223         .kr_pcieq0serdesstatus = IPATH_KREG_OFFSET(PCIEQ0SerdesStatus),
 224         .kr_pcieq1serdesconfig0 = IPATH_KREG_OFFSET(PCIEQ1SerdesConfig0),
 225         .kr_pcieq1serdesconfig1 = IPATH_KREG_OFFSET(PCIEQ1SerdesConfig1),
 226         .kr_pcieq1serdesstatus = IPATH_KREG_OFFSET(PCIEQ1SerdesStatus)
 227 };
 228
 229 static const struct ipath_cregs ipath_pe_cregs = {
 230         .cr_badformatcnt = IPATH_CREG_OFFSET(RxBadFormatCnt),
 231         .cr_erricrccnt = IPATH_CREG_OFFSET(RxICRCErrCnt),
 232         .cr_errlinkcnt = IPATH_CREG_OFFSET(RxLinkProblemCnt),
 233         .cr_errlpcrccnt = IPATH_CREG_OFFSET(RxLPCRCErrCnt),
 234         .cr_errpkey = IPATH_CREG_OFFSET(RxPKeyMismatchCnt),
 235         .cr_errrcvflowctrlcnt = IPATH_CREG_OFFSET(RxFlowCtrlErrCnt),
 236         .cr_err_rlencnt = IPATH_CREG_OFFSET(RxLenErrCnt),
 237         .cr_errslencnt = IPATH_CREG_OFFSET(TxLenErrCnt),
 238         .cr_errtidfull = IPATH_CREG_OFFSET(RxTIDFullErrCnt),
 239         .cr_errtidvalid = IPATH_CREG_OFFSET(RxTIDValidErrCnt),
 240         .cr_errvcrccnt = IPATH_CREG_OFFSET(RxVCRCErrCnt),
 241         .cr_ibstatuschange = IPATH_CREG_OFFSET(IBStatusChangeCnt),
 242         .cr_intcnt = IPATH_CREG_OFFSET(LBIntCnt),
 243         .cr_invalidrlencnt = IPATH_CREG_OFFSET(RxMaxMinLenErrCnt),
 244         .cr_invalidslencnt = IPATH_CREG_OFFSET(TxMaxMinLenErrCnt),
 245         .cr_lbflowstallcnt = IPATH_CREG_OFFSET(LBFlowStallCnt),
 246         .cr_pktrcvcnt = IPATH_CREG_OFFSET(RxDataPktCnt),
 247         .cr_pktrcvflowctrlcnt = IPATH_CREG_OFFSET(RxFlowPktCnt),
 248         .cr_pktsendcnt = IPATH_CREG_OFFSET(TxDataPktCnt),
 249         .cr_pktsendflowcnt = IPATH_CREG_OFFSET(TxFlowPktCnt),
 250         .cr_portovflcnt = IPATH_CREG_OFFSET(RxP0HdrEgrOvflCnt),
 251         .cr_rcvebpcnt = IPATH_CREG_OFFSET(RxEBPCnt),
 252         .cr_rcvovflcnt = IPATH_CREG_OFFSET(RxBufOvflCnt),
 253         .cr_senddropped = IPATH_CREG_OFFSET(TxDroppedPktCnt),
 254         .cr_sendstallcnt = IPATH_CREG_OFFSET(TxFlowStallCnt),
 255         .cr_sendunderruncnt = IPATH_CREG_OFFSET(TxUnderrunCnt),
 256         .cr_wordrcvcnt = IPATH_CREG_OFFSET(RxDwordCnt),
 257         .cr_wordsendcnt = IPATH_CREG_OFFSET(TxDwordCnt),
 258         .cr_unsupvlcnt = IPATH_CREG_OFFSET(TxUnsupVLErrCnt),
 259         .cr_rxdroppktcnt = IPATH_CREG_OFFSET(RxDroppedPktCnt),
 260         .cr_iblinkerrrecovcnt = IPATH_CREG_OFFSET(IBLinkErrRecoveryCnt),
 261         .cr_iblinkdowncnt = IPATH_CREG_OFFSET(IBLinkDownedCnt),
 262         .cr_ibsymbolerrcnt = IPATH_CREG_OFFSET(IBSymbolErrCnt)
 263 };
 264
 265 /* kr_intstatus, kr_intclear, kr_intmask bits */
 266 #define INFINIPATH_I_RCVURG_MASK ((1U<<5)-1)
 267 #define INFINIPATH_I_RCVAVAIL_MASK ((1U<<5)-1)
 268
 269 /* kr_hwerrclear, kr_hwerrmask, kr_hwerrstatus, bits */
 270 #define INFINIPATH_HWE_PCIEMEMPARITYERR_MASK  0x000000000000003fULL
 271 #define INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT 0
 272 #define INFINIPATH_HWE_PCIEPOISONEDTLP      0x0000000010000000ULL
 273 #define INFINIPATH_HWE_PCIECPLTIMEOUT       0x0000000020000000ULL
 274 #define INFINIPATH_HWE_PCIEBUSPARITYXTLH    0x0000000040000000ULL
 275 #define INFINIPATH_HWE_PCIEBUSPARITYXADM    0x0000000080000000ULL
 276 #define INFINIPATH_HWE_PCIEBUSPARITYRADM    0x0000000100000000ULL
 277 #define INFINIPATH_HWE_COREPLL_FBSLIP       0x0080000000000000ULL
 278 #define INFINIPATH_HWE_COREPLL_RFSLIP       0x0100000000000000ULL
 279 #define INFINIPATH_HWE_PCIE1PLLFAILED       0x0400000000000000ULL
 280 #define INFINIPATH_HWE_PCIE0PLLFAILED       0x0800000000000000ULL
 281 #define INFINIPATH_HWE_SERDESPLLFAILED      0x1000000000000000ULL
 282
 283 /* kr_extstatus bits */
 284 #define INFINIPATH_EXTS_FREQSEL 0x2
 285 #define INFINIPATH_EXTS_SERDESSEL 0x4
 286 #define INFINIPATH_EXTS_MEMBIST_ENDTEST     0x0000000000004000
 287 #define INFINIPATH_EXTS_MEMBIST_FOUND       0x0000000000008000
 288
 289 #define _IPATH_GPIO_SDA_NUM 1
 290 #define _IPATH_GPIO_SCL_NUM 0
 291
 292 #define IPATH_GPIO_SDA (1ULL << \
 293         (_IPATH_GPIO_SDA_NUM+INFINIPATH_EXTC_GPIOOE_SHIFT))
 294 #define IPATH_GPIO_SCL (1ULL << \
 295         (_IPATH_GPIO_SCL_NUM+INFINIPATH_EXTC_GPIOOE_SHIFT))
 296
 297 /*
 298  * Rev2 silicon allows suppressing check for ArmLaunch errors.
 299  * this can speed up short packet sends on systems that do
 300  * not guaranteee write-order.
 301  */
 302 #define INFINIPATH_XGXS_SUPPRESS_ARMLAUNCH_ERR (1ULL<<63)
 303
 304 /* 6120 specific hardware errors... */
 305 static const struct ipath_hwerror_msgs ipath_6120_hwerror_msgs[] = {
 306         INFINIPATH_HWE_MSG(PCIEPOISONEDTLP, "PCIe Poisoned TLP"),
 307         INFINIPATH_HWE_MSG(PCIECPLTIMEOUT, "PCIe completion timeout"),
 308         /*
 309          * In practice, it's unlikely wthat we'll see PCIe PLL, or bus
 310          * parity or memory parity error failures, because most likely we
 311          * won't be able to talk to the core of the chip.  Nonetheless, we
 312          * might see them, if they are in parts of the PCIe core that aren't
 313          * essential.
 314          */
 315         INFINIPATH_HWE_MSG(PCIE1PLLFAILED, "PCIePLL1"),
 316         INFINIPATH_HWE_MSG(PCIE0PLLFAILED, "PCIePLL0"),
 317         INFINIPATH_HWE_MSG(PCIEBUSPARITYXTLH, "PCIe XTLH core parity"),
 318         INFINIPATH_HWE_MSG(PCIEBUSPARITYXADM, "PCIe ADM TX core parity"),
 319         INFINIPATH_HWE_MSG(PCIEBUSPARITYRADM, "PCIe ADM RX core parity"),
 320         INFINIPATH_HWE_MSG(RXDSYNCMEMPARITYERR, "Rx Dsync"),
 321         INFINIPATH_HWE_MSG(SERDESPLLFAILED, "SerDes PLL"),
 322 };
 323
 324 /**
 325  * ipath_pe_handle_hwerrors - display hardware errors.
 326  * @dd: the infinipath device
 327  * @msg: the output buffer
 328  * @msgl: the size of the output buffer
 329  *
 330  * Use same msg buffer as regular errors to avoid excessive stack
 331  * use.  Most hardware errors are catastrophic, but for right now,
 332  * we'll print them and continue.  We reuse the same message buffer as
 333  * ipath_handle_errors() to avoid excessive stack usage.
 334  */
 335 static void ipath_pe_handle_hwerrors(struct ipath_devdata *dd, char *msg,
 336                                      size_t msgl)
 337 {
 338         ipath_err_t hwerrs;
 339         u32 bits, ctrl;
 340         int isfatal = 0;
 341         char bitsmsg[64];
 342
 343         hwerrs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_hwerrstatus);
 344         if (!hwerrs) {
 345                 /*
 346                  * better than printing cofusing messages
 347                  * This seems to be related to clearing the crc error, or
 348                  * the pll error during init.
 349                  */
 350                 ipath_cdbg(VERBOSE, "Called but no hardware errors set\n");
 351                 return;
 352         } else if (hwerrs == ~0ULL) {
 353                 ipath_dev_err(dd, "Read of hardware error status failed "
 354                               "(all bits set); ignoring\n");
 355                 return;
 356         }
 357         ipath_stats.sps_hwerrs++;
 358
 359         /* Always clear the error status register, except MEMBISTFAIL,
 360          * regardless of whether we continue or stop using the chip.
 361          * We want that set so we know it failed, even across driver reload.
 362          * We'll still ignore it in the hwerrmask.  We do this partly for
 363          * diagnostics, but also for support */
 364         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 365                          hwerrs&~INFINIPATH_HWE_MEMBISTFAILED);
 366
 367         hwerrs &= dd->ipath_hwerrmask;
 368
 369         /*
 370          * make sure we get this much out, unless told to be quiet,
 371          * or it's occurred within the last 5 seconds
 372          */
 373         if ((hwerrs & ~(dd->ipath_lasthwerror |
 374                         ((INFINIPATH_HWE_TXEMEMPARITYERR_PIOBUF |
 375                           INFINIPATH_HWE_TXEMEMPARITYERR_PIOPBC)
 376                          << INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT))) ||
 377             (ipath_debug & __IPATH_VERBDBG))
 378                 dev_info(&dd->pcidev->dev, "Hardware error: hwerr=0x%llx "
 379                          "(cleared)\n", (unsigned long long) hwerrs);
 380         dd->ipath_lasthwerror |= hwerrs;
 381
 382         if (hwerrs & ~dd->ipath_hwe_bitsextant)
 383                 ipath_dev_err(dd, "hwerror interrupt with unknown errors "
 384                               "%llx set\n", (unsigned long long)
 385                               (hwerrs & ~dd->ipath_hwe_bitsextant));
 386
 387         ctrl = ipath_read_kreg32(dd, dd->ipath_kregs->kr_control);
 388         if (ctrl & INFINIPATH_C_FREEZEMODE) {
 389                 /*
 390                  * parity errors in send memory are recoverable,
 391                  * just cancel the send (if indicated in * sendbuffererror),
 392                  * count the occurrence, unfreeze (if no other handled
 393                  * hardware error bits are set), and continue. They can
 394                  * occur if a processor speculative read is done to the PIO
 395                  * buffer while we are sending a packet, for example.
 396                  */
 397                 if (hwerrs & ((INFINIPATH_HWE_TXEMEMPARITYERR_PIOBUF |
 398                                INFINIPATH_HWE_TXEMEMPARITYERR_PIOPBC)
 399                               << INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT)) {
 400                         ipath_stats.sps_txeparity++;
 401                         ipath_dbg("Recovering from TXE parity error (%llu), "
 402                                   "hwerrstatus=%llx\n",
 403                                   (unsigned long long) ipath_stats.sps_txeparity,
 404                                   (unsigned long long) hwerrs);
 405                         ipath_disarm_senderrbufs(dd);
 406                         hwerrs &= ~((INFINIPATH_HWE_TXEMEMPARITYERR_PIOBUF |
 407                                      INFINIPATH_HWE_TXEMEMPARITYERR_PIOPBC)
 408                                     << INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT);
 409                         if (!hwerrs) { /* else leave in freeze mode */
 410                                 ipath_write_kreg(dd,
 411                                                  dd->ipath_kregs->kr_control,
 412                                                  dd->ipath_control);
 413                             return;
 414                         }
 415                 }
 416                 if (hwerrs) {
 417                         /*
 418                          * if any set that we aren't ignoring only make the
 419                          * complaint once, in case it's stuck or recurring,
 420                          * and we get here multiple times
 421                          */
 422                         if (dd->ipath_flags & IPATH_INITTED) {
 423                                 ipath_dev_err(dd, "Fatal Hardware Error (freeze "
 424                                               "mode), no longer usable, SN %.16s\n",
 425                                                   dd->ipath_serial);
 426                                 isfatal = 1;
 427                         }
 428                         /*
 429                          * Mark as having had an error for driver, and also
 430                          * for /sys and status word mapped to user programs.
 431                          * This marks unit as not usable, until reset
 432                          */
 433                         *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
 434                         *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
 435                         dd->ipath_flags &= ~IPATH_INITTED;
 436                 } else {
 437                         ipath_dbg("Clearing freezemode on ignored hardware "
 438                                   "error\n");
 439                         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
 440                                          dd->ipath_control);
 441                 }
 442         }
 443
 444         *msg = '\0';
 445
 446         if (hwerrs & INFINIPATH_HWE_MEMBISTFAILED) {
 447                 strlcat(msg, "[Memory BIST test failed, InfiniPath hardware unusable]",
 448                         msgl);
 449                 /* ignore from now on, so disable until driver reloaded */
 450                 *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
 451                 dd->ipath_hwerrmask &= ~INFINIPATH_HWE_MEMBISTFAILED;
 452                 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
 453                                  dd->ipath_hwerrmask);
 454         }
 455
 456         ipath_format_hwerrors(hwerrs,
 457                               ipath_6120_hwerror_msgs,
 458                               sizeof(ipath_6120_hwerror_msgs)/
 459                               sizeof(ipath_6120_hwerror_msgs[0]),
 460                               msg, msgl);
 461
 462         if (hwerrs & (INFINIPATH_HWE_PCIEMEMPARITYERR_MASK
 463                       << INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT)) {
 464                 bits = (u32) ((hwerrs >>
 465                                INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT) &
 466                               INFINIPATH_HWE_PCIEMEMPARITYERR_MASK);
 467                 snprintf(bitsmsg, sizeof bitsmsg,
 468                          "[PCIe Mem Parity Errs %x] ", bits);
 469                 strlcat(msg, bitsmsg, msgl);
 470         }
 471
 472 #define _IPATH_PLL_FAIL (INFINIPATH_HWE_COREPLL_FBSLIP |        \
 473                          INFINIPATH_HWE_COREPLL_RFSLIP )
 474
 475         if (hwerrs & _IPATH_PLL_FAIL) {
 476                 snprintf(bitsmsg, sizeof bitsmsg,
 477                          "[PLL failed (%llx), InfiniPath hardware unusable]",
 478                          (unsigned long long) hwerrs & _IPATH_PLL_FAIL);
 479                 strlcat(msg, bitsmsg, msgl);
 480                 /* ignore from now on, so disable until driver reloaded */
 481                 dd->ipath_hwerrmask &= ~(hwerrs & _IPATH_PLL_FAIL);
 482                 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
 483                                  dd->ipath_hwerrmask);
 484         }
 485
 486         if (hwerrs & INFINIPATH_HWE_SERDESPLLFAILED) {
 487                 /*
 488                  * If it occurs, it is left masked since the eternal
 489                  * interface is unused
 490                  */
 491                 dd->ipath_hwerrmask &= ~INFINIPATH_HWE_SERDESPLLFAILED;
 492                 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
 493                                  dd->ipath_hwerrmask);
 494         }
 495
 496         ipath_dev_err(dd, "%s hardware error\n", msg);
 497         if (isfatal && !ipath_diag_inuse && dd->ipath_freezemsg) {
 498                 /*
 499                  * for /sys status file ; if no trailing } is copied, we'll
 500                  * know it was truncated.
 501                  */
 502                 snprintf(dd->ipath_freezemsg, dd->ipath_freezelen,
 503                          "{%s}", msg);
 504         }
 505 }
 506
 507 /**
 508  * ipath_pe_boardname - fill in the board name
 509  * @dd: the infinipath device
 510  * @name: the output buffer
 511  * @namelen: the size of the output buffer
 512  *
 513  * info is based on the board revision register
 514  */
 515 static int ipath_pe_boardname(struct ipath_devdata *dd, char *name,
 516                               size_t namelen)
 517 {
 518         char *n = NULL;
 519         u8 boardrev = dd->ipath_boardrev;
 520         int ret;
 521
 522         switch (boardrev) {
 523         case 0:
 524                 n = "InfiniPath_Emulation";
 525                 break;
 526         case 1:
 527                 n = "InfiniPath_QLE7140-Bringup";
 528                 break;
 529         case 2:
 530                 n = "InfiniPath_QLE7140";
 531                 break;
 532         case 3:
 533                 n = "InfiniPath_QMI7140";
 534                 break;
 535         case 4:
 536                 n = "InfiniPath_QEM7140";
 537                 break;
 538         case 5:
 539                 n = "InfiniPath_QMH7140";
 540                 break;
 541         case 6:
 542                 n = "InfiniPath_QLE7142";
 543                 break;
 544         default:
 545                 ipath_dev_err(dd,
 546                               "Don't yet know about board with ID %u\n",
 547                               boardrev);
 548                 snprintf(name, namelen, "Unknown_InfiniPath_PCIe_%u",
 549                          boardrev);
 550                 break;
 551         }
 552         if (n)
 553                 snprintf(name, namelen, "%s", n);
 554
 555         if (dd->ipath_majrev != 4 || !dd->ipath_minrev || dd->ipath_minrev>2) {
 556                 ipath_dev_err(dd, "Unsupported InfiniPath hardware revision %u.%u!\n",
 557                               dd->ipath_majrev, dd->ipath_minrev);
 558                 ret = 1;
 559         } else
 560                 ret = 0;
 561
 562         return ret;
 563 }
 564
 565 /**
 566  * ipath_pe_init_hwerrors - enable hardware errors
 567  * @dd: the infinipath device
 568  *
 569  * now that we have finished initializing everything that might reasonably
 570  * cause a hardware error, and cleared those errors bits as they occur,
 571  * we can enable hardware errors in the mask (potentially enabling
 572  * freeze mode), and enable hardware errors as errors (along with
 573  * everything else) in errormask
 574  */
 575 static void ipath_pe_init_hwerrors(struct ipath_devdata *dd)
 576 {
 577         ipath_err_t val;
 578         u64 extsval;
 579
 580         extsval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_extstatus);
 581
 582         if (!(extsval & INFINIPATH_EXTS_MEMBIST_ENDTEST))
 583                 ipath_dev_err(dd, "MemBIST did not complete!\n");
 584
 585         val = ~0ULL;    /* barring bugs, all hwerrors become interrupts, */
 586
 587         if (!dd->ipath_boardrev)        // no PLL for Emulator
 588                 val &= ~INFINIPATH_HWE_SERDESPLLFAILED;
 589
 590         if (dd->ipath_minrev < 2) {
 591                 /* workaround bug 9460 in internal interface bus parity
 592                  * checking. Fixed (HW bug 9490) in Rev2.
 593                  */
 594                 val &= ~INFINIPATH_HWE_PCIEBUSPARITYRADM;
 595         }
 596         dd->ipath_hwerrmask = val;
 597 }
 598
 599 /**
 600  * ipath_pe_bringup_serdes - bring up the serdes
 601  * @dd: the infinipath device
 602  */
 603 static int ipath_pe_bringup_serdes(struct ipath_devdata *dd)
 604 {
 605         u64 val, tmp, config1, prev_val;
 606         int ret = 0;
 607
 608         ipath_dbg("Trying to bringup serdes\n");
 609
 610         if (ipath_read_kreg64(dd, dd->ipath_kregs->kr_hwerrstatus) &
 611             INFINIPATH_HWE_SERDESPLLFAILED) {
 612                 ipath_dbg("At start, serdes PLL failed bit set "
 613                           "in hwerrstatus, clearing and continuing\n");
 614                 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
 615                                  INFINIPATH_HWE_SERDESPLLFAILED);
 616         }
 617
 618         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig0);
 619         config1 = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig1);
 620
 621         ipath_cdbg(VERBOSE, "SerDes status config0=%llx config1=%llx, "
 622                    "xgxsconfig %llx\n", (unsigned long long) val,
 623                    (unsigned long long) config1, (unsigned long long)
 624                    ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig));
 625
 626         /*
 627          * Force reset on, also set rxdetect enable.  Must do before reading
 628          * serdesstatus at least for simulation, or some of the bits in
 629          * serdes status will come back as undefined and cause simulation
 630          * failures
 631          */
 632         val |= INFINIPATH_SERDC0_RESET_PLL | INFINIPATH_SERDC0_RXDETECT_EN
 633                 | INFINIPATH_SERDC0_L1PWR_DN;
 634         ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
 635         /* be sure chip saw it */
 636         tmp = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 637         udelay(5);              /* need pll reset set at least for a bit */
 638         /*
 639          * after PLL is reset, set the per-lane Resets and TxIdle and
 640          * clear the PLL reset and rxdetect (to get falling edge).
 641          * Leave L1PWR bits set (permanently)
 642          */
 643         val &= ~(INFINIPATH_SERDC0_RXDETECT_EN | INFINIPATH_SERDC0_RESET_PLL
 644                  | INFINIPATH_SERDC0_L1PWR_DN);
 645         val |= INFINIPATH_SERDC0_RESET_MASK | INFINIPATH_SERDC0_TXIDLE;
 646         ipath_cdbg(VERBOSE, "Clearing pll reset and setting lane resets "
 647                    "and txidle (%llx)\n", (unsigned long long) val);
 648         ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
 649         /* be sure chip saw it */
 650         tmp = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 651         /* need PLL reset clear for at least 11 usec before lane
 652          * resets cleared; give it a few more to be sure */
 653         udelay(15);
 654         val &= ~(INFINIPATH_SERDC0_RESET_MASK | INFINIPATH_SERDC0_TXIDLE);
 655
 656         ipath_cdbg(VERBOSE, "Clearing lane resets and txidle "
 657                    "(writing %llx)\n", (unsigned long long) val);
 658         ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
 659         /* be sure chip saw it */
 660         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 661
 662         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig);
 663         prev_val = val;
 664         if (((val >> INFINIPATH_XGXS_MDIOADDR_SHIFT) &
 665              INFINIPATH_XGXS_MDIOADDR_MASK) != 3) {
 666                 val &=
 667                         ~(INFINIPATH_XGXS_MDIOADDR_MASK <<
 668                           INFINIPATH_XGXS_MDIOADDR_SHIFT);
 669                 /* MDIO address 3 */
 670                 val |= 3ULL << INFINIPATH_XGXS_MDIOADDR_SHIFT;
 671         }
 672         if (val & INFINIPATH_XGXS_RESET) {
 673                 val &= ~INFINIPATH_XGXS_RESET;
 674         }
 675         if (((val >> INFINIPATH_XGXS_RX_POL_SHIFT) &
 676              INFINIPATH_XGXS_RX_POL_MASK) != dd->ipath_rx_pol_inv ) {
 677                 /* need to compensate for Tx inversion in partner */
 678                 val &= ~(INFINIPATH_XGXS_RX_POL_MASK <<
 679                          INFINIPATH_XGXS_RX_POL_SHIFT);
 680                 val |= dd->ipath_rx_pol_inv <<
 681                         INFINIPATH_XGXS_RX_POL_SHIFT;
 682         }
 683         if (dd->ipath_minrev >= 2) {
 684                 /* Rev 2. can tolerate multiple writes to PBC, and
 685                  * allowing them can provide lower latency on some
 686                  * CPUs, but this feature is off by default, only
 687                  * turned on by setting D63 of XGXSconfig reg.
 688                  * May want to make this conditional more
 689                  * fine-grained in future. This is not exactly
 690                  * related to XGXS, but where the bit ended up.
 691                  */
 692                 val |= INFINIPATH_XGXS_SUPPRESS_ARMLAUNCH_ERR;
 693         }
 694         if (val != prev_val)
 695                 ipath_write_kreg(dd, dd->ipath_kregs->kr_xgxsconfig, val);
 696
 697         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig0);
 698
 699         /* clear current and de-emphasis bits */
 700         config1 &= ~0x0ffffffff00ULL;
 701         /* set current to 20ma */
 702         config1 |= 0x00000000000ULL;
 703         /* set de-emphasis to -5.68dB */
 704         config1 |= 0x0cccc000000ULL;
 705         ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig1, config1);
 706
 707         ipath_cdbg(VERBOSE, "done: SerDes status config0=%llx "
 708                    "config1=%llx, sstatus=%llx xgxs=%llx\n",
 709                    (unsigned long long) val, (unsigned long long) config1,
 710                    (unsigned long long)
 711                    ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesstatus),
 712                    (unsigned long long)
 713                    ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig));
 714
 715         if (!ipath_waitfor_mdio_cmdready(dd)) {
 716                 ipath_write_kreg(
 717                         dd, dd->ipath_kregs->kr_mdio,
 718                         ipath_mdio_req(IPATH_MDIO_CMD_READ, 31,
 719                                        IPATH_MDIO_CTRL_XGXS_REG_8, 0));
 720                 if (ipath_waitfor_complete(dd, dd->ipath_kregs->kr_mdio,
 721                                            IPATH_MDIO_DATAVALID, &val))
 722                         ipath_dbg("Never got MDIO data for XGXS "
 723                                   "status read\n");
 724                 else
 725                         ipath_cdbg(VERBOSE, "MDIO Read reg8, "
 726                                    "'bank' 31 %x\n", (u32) val);
 727         } else
 728                 ipath_dbg("Never got MDIO cmdready for XGXS status read\n");
 729
 730         return ret;
 731 }
 732
 733 /**
 734  * ipath_pe_quiet_serdes - set serdes to txidle
 735  * @dd: the infinipath device
 736  * Called when driver is being unloaded
 737  */
 738 static void ipath_pe_quiet_serdes(struct ipath_devdata *dd)
 739 {
 740         u64 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig0);
 741
 742         val |= INFINIPATH_SERDC0_TXIDLE;
 743         ipath_dbg("Setting TxIdleEn on serdes (config0 = %llx)\n",
 744                   (unsigned long long) val);
 745         ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
 746 }
 747
 748 static int ipath_pe_intconfig(struct ipath_devdata *dd)
 749 {
 750         u64 val;
 751         u32 chiprev;
 752
 753         /*
 754          * If the chip supports added error indication via GPIO pins,
 755          * enable interrupts on those bits so the interrupt routine
 756          * can count the events. Also set flag so interrupt routine
 757          * can know they are expected.
 758          */
 759         chiprev = dd->ipath_revision >> INFINIPATH_R_CHIPREVMINOR_SHIFT;
 760         if ((chiprev & INFINIPATH_R_CHIPREVMINOR_MASK) > 1) {
 761                 /* Rev2+ reports extra errors via internal GPIO pins */
 762                 dd->ipath_flags |= IPATH_GPIO_ERRINTRS;
 763                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_gpio_mask);
 764                 val |= IPATH_GPIO_ERRINTR_MASK;
 765                 ipath_write_kreg( dd, dd->ipath_kregs->kr_gpio_mask, val);
 766         }
 767         return 0;
 768 }
 769
 770 /**
 771  * ipath_setup_pe_setextled - set the state of the two external LEDs
 772  * @dd: the infinipath device
 773  * @lst: the L state
 774  * @ltst: the LT state
 775
 776  * These LEDs indicate the physical and logical state of IB link.
 777  * For this chip (at least with recommended board pinouts), LED1
 778  * is Yellow (logical state) and LED2 is Green (physical state),
 779  *
 780  * Note:  We try to match the Mellanox HCA LED behavior as best
 781  * we can.  Green indicates physical link state is OK (something is
 782  * plugged in, and we can train).
 783  * Amber indicates the link is logically up (ACTIVE).
 784  * Mellanox further blinks the amber LED to indicate data packet
 785  * activity, but we have no hardware support for that, so it would
 786  * require waking up every 10-20 msecs and checking the counters
 787  * on the chip, and then turning the LED off if appropriate.  That's
 788  * visible overhead, so not something we will do.
 789  *
 790  */
 791 static void ipath_setup_pe_setextled(struct ipath_devdata *dd, u64 lst,
 792                                      u64 ltst)
 793 {
 794         u64 extctl;
 795
 796         /* the diags use the LED to indicate diag info, so we leave
 797          * the external LED alone when the diags are running */
 798         if (ipath_diag_inuse)
 799                 return;
 800
 801         extctl = dd->ipath_extctrl & ~(INFINIPATH_EXTC_LED1PRIPORT_ON |
 802                                        INFINIPATH_EXTC_LED2PRIPORT_ON);
 803
 804         if (ltst & INFINIPATH_IBCS_LT_STATE_LINKUP)
 805                 extctl |= INFINIPATH_EXTC_LED2PRIPORT_ON;
 806         if (lst == INFINIPATH_IBCS_L_STATE_ACTIVE)
 807                 extctl |= INFINIPATH_EXTC_LED1PRIPORT_ON;
 808         dd->ipath_extctrl = extctl;
 809         ipath_write_kreg(dd, dd->ipath_kregs->kr_extctrl, extctl);
 810 }
 811
 812 /**
 813  * ipath_setup_pe_cleanup - clean up any per-chip chip-specific stuff
 814  * @dd: the infinipath device
 815  *
 816  * This is called during driver unload.
 817  * We do the pci_disable_msi here, not in generic code, because it
 818  * isn't used for the HT chips. If we do end up needing pci_enable_msi
 819  * at some point in the future for HT, we'll move the call back
 820  * into the main init_one code.
 821  */
 822 static void ipath_setup_pe_cleanup(struct ipath_devdata *dd)
 823 {
 824         dd->ipath_msi_lo = 0;   /* just in case unload fails */
 825         pci_disable_msi(dd->pcidev);
 826 }
 827
 828 /**
 829  * ipath_setup_pe_config - setup PCIe config related stuff
 830  * @dd: the infinipath device
 831  * @pdev: the PCI device
 832  *
 833  * The pci_enable_msi() call will fail on systems with MSI quirks
 834  * such as those with AMD8131, even if the device of interest is not
 835  * attached to that device, (in the 2.6.13 - 2.6.15 kernels, at least, fixed
 836  * late in 2.6.16).
 837  * All that can be done is to edit the kernel source to remove the quirk
 838  * check until that is fixed.
 839  * We do not need to call enable_msi() for our HyperTransport chip,
 840  * even though it uses MSI, and we want to avoid the quirk warning, so
 841  * So we call enable_msi only for PCIe.  If we do end up needing
 842  * pci_enable_msi at some point in the future for HT, we'll move the
 843  * call back into the main init_one code.
 844  * We save the msi lo and hi values, so we can restore them after
 845  * chip reset (the kernel PCI infrastructure doesn't yet handle that
 846  * correctly).
 847  */
 848 static int ipath_setup_pe_config(struct ipath_devdata *dd,
 849                                  struct pci_dev *pdev)
 850 {
 851         int pos, ret;
 852
 853         dd->ipath_msi_lo = 0;   /* used as a flag during reset processing */
 854         ret = pci_enable_msi(dd->pcidev);
 855         if (ret)
 856                 ipath_dev_err(dd, "pci_enable_msi failed: %d, "
 857                               "interrupts may not work\n", ret);
 858         /* continue even if it fails, we may still be OK... */
 859
 860         if ((pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI))) {
 861                 u16 control;
 862                 pci_read_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
 863                                       &dd->ipath_msi_lo);
 864                 pci_read_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
 865                                       &dd->ipath_msi_hi);
 866                 pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
 867                                      &control);
 868                 /* now save the data (vector) info */
 869                 pci_read_config_word(dd->pcidev,
 870                                      pos + ((control & PCI_MSI_FLAGS_64BIT)
 871                                             ? 12 : 8),
 872                                      &dd->ipath_msi_data);
 873                 ipath_cdbg(VERBOSE, "Read msi data 0x%x from config offset "
 874                            "0x%x, control=0x%x\n", dd->ipath_msi_data,
 875                            pos + ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
 876                            control);
 877                 /* we save the cachelinesize also, although it doesn't
 878                  * really matter */
 879                 pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE,
 880                                      &dd->ipath_pci_cacheline);
 881         } else
 882                 ipath_dev_err(dd, "Can't find MSI capability, "
 883                               "can't save MSI settings for reset\n");
 884         if ((pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_EXP))) {
 885                 u16 linkstat;
 886                 pci_read_config_word(dd->pcidev, pos + PCI_EXP_LNKSTA,
 887                                      &linkstat);
 888                 linkstat >>= 4;
 889                 linkstat &= 0x1f;
 890                 if (linkstat != 8)
 891                         ipath_dev_err(dd, "PCIe width %u, "
 892                                       "performance reduced\n", linkstat);
 893         }
 894         else
 895                 ipath_dev_err(dd, "Can't find PCI Express "
 896                               "capability!\n");
 897         return 0;
 898 }
 899
 900 static void ipath_init_pe_variables(struct ipath_devdata *dd)
 901 {
 902         /*
 903          * bits for selecting i2c direction and values,
 904          * used for I2C serial flash
 905          */
 906         dd->ipath_gpio_sda_num = _IPATH_GPIO_SDA_NUM;
 907         dd->ipath_gpio_scl_num = _IPATH_GPIO_SCL_NUM;
 908         dd->ipath_gpio_sda = IPATH_GPIO_SDA;
 909         dd->ipath_gpio_scl = IPATH_GPIO_SCL;
 910
 911         /* variables for sanity checking interrupt and errors */
 912         dd->ipath_hwe_bitsextant =
 913                 (INFINIPATH_HWE_RXEMEMPARITYERR_MASK <<
 914                  INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT) |
 915                 (INFINIPATH_HWE_TXEMEMPARITYERR_MASK <<
 916                  INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT) |
 917                 (INFINIPATH_HWE_PCIEMEMPARITYERR_MASK <<
 918                  INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT) |
 919                 INFINIPATH_HWE_PCIE1PLLFAILED |
 920                 INFINIPATH_HWE_PCIE0PLLFAILED |
 921                 INFINIPATH_HWE_PCIEPOISONEDTLP |
 922                 INFINIPATH_HWE_PCIECPLTIMEOUT |
 923                 INFINIPATH_HWE_PCIEBUSPARITYXTLH |
 924                 INFINIPATH_HWE_PCIEBUSPARITYXADM |
 925                 INFINIPATH_HWE_PCIEBUSPARITYRADM |
 926                 INFINIPATH_HWE_MEMBISTFAILED |
 927                 INFINIPATH_HWE_COREPLL_FBSLIP |
 928                 INFINIPATH_HWE_COREPLL_RFSLIP |
 929                 INFINIPATH_HWE_SERDESPLLFAILED |
 930                 INFINIPATH_HWE_IBCBUSTOSPCPARITYERR |
 931                 INFINIPATH_HWE_IBCBUSFRSPCPARITYERR;
 932         dd->ipath_i_bitsextant =
 933                 (INFINIPATH_I_RCVURG_MASK << INFINIPATH_I_RCVURG_SHIFT) |
 934                 (INFINIPATH_I_RCVAVAIL_MASK <<
 935                  INFINIPATH_I_RCVAVAIL_SHIFT) |
 936                 INFINIPATH_I_ERROR | INFINIPATH_I_SPIOSENT |
 937                 INFINIPATH_I_SPIOBUFAVAIL | INFINIPATH_I_GPIO;
 938         dd->ipath_e_bitsextant =
 939                 INFINIPATH_E_RFORMATERR | INFINIPATH_E_RVCRC |
 940                 INFINIPATH_E_RICRC | INFINIPATH_E_RMINPKTLEN |
 941                 INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RLONGPKTLEN |
 942                 INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RUNEXPCHAR |
 943                 INFINIPATH_E_RUNSUPVL | INFINIPATH_E_REBP |
 944                 INFINIPATH_E_RIBFLOW | INFINIPATH_E_RBADVERSION |
 945                 INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
 946                 INFINIPATH_E_RBADTID | INFINIPATH_E_RHDRLEN |
 947                 INFINIPATH_E_RHDR | INFINIPATH_E_RIBLOSTLINK |
 948                 INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SMAXPKTLEN |
 949                 INFINIPATH_E_SUNDERRUN | INFINIPATH_E_SPKTLEN |
 950                 INFINIPATH_E_SDROPPEDSMPPKT | INFINIPATH_E_SDROPPEDDATAPKT |
 951                 INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM |
 952                 INFINIPATH_E_SUNSUPVL | INFINIPATH_E_IBSTATUSCHANGED |
 953                 INFINIPATH_E_INVALIDADDR | INFINIPATH_E_RESET |
 954                 INFINIPATH_E_HARDWARE;
 955
 956         dd->ipath_i_rcvavail_mask = INFINIPATH_I_RCVAVAIL_MASK;
 957         dd->ipath_i_rcvurg_mask = INFINIPATH_I_RCVURG_MASK;
 958 }
 959
 960 /* setup the MSI stuff again after a reset.  I'd like to just call
 961  * pci_enable_msi() and request_irq() again, but when I do that,
 962  * the MSI enable bit doesn't get set in the command word, and
 963  * we switch to to a different interrupt vector, which is confusing,
 964  * so I instead just do it all inline.  Perhaps somehow can tie this
 965  * into the PCIe hotplug support at some point
 966  * Note, because I'm doing it all here, I don't call pci_disable_msi()
 967  * or free_irq() at the start of ipath_setup_pe_reset().
 968  */
 969 static int ipath_reinit_msi(struct ipath_devdata *dd)
 970 {
 971         int pos;
 972         u16 control;
 973         int ret;
 974
 975         if (!dd->ipath_msi_lo) {
 976                 dev_info(&dd->pcidev->dev, "Can't restore MSI config, "
 977                          "initial setup failed?\n");
 978                 ret = 0;
 979                 goto bail;
 980         }
 981
 982         if (!(pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI))) {
 983                 ipath_dev_err(dd, "Can't find MSI capability, "
 984                               "can't restore MSI settings\n");
 985                 ret = 0;
 986                 goto bail;
 987         }
 988         ipath_cdbg(VERBOSE, "Writing msi_lo 0x%x to config offset 0x%x\n",
 989                    dd->ipath_msi_lo, pos + PCI_MSI_ADDRESS_LO);
 990         pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
 991                                dd->ipath_msi_lo);
 992         ipath_cdbg(VERBOSE, "Writing msi_lo 0x%x to config offset 0x%x\n",
 993                    dd->ipath_msi_hi, pos + PCI_MSI_ADDRESS_HI);
 994         pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
 995                                dd->ipath_msi_hi);
 996         pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
 997         if (!(control & PCI_MSI_FLAGS_ENABLE)) {
 998                 ipath_cdbg(VERBOSE, "MSI control at off %x was %x, "
 999                            "setting MSI enable (%x)\n", pos + PCI_MSI_FLAGS,
1000                            control, control | PCI_MSI_FLAGS_ENABLE);
1001                 control |= PCI_MSI_FLAGS_ENABLE;
1002                 pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
1003                                       control);
1004         }
1005         /* now rewrite the data (vector) info */
1006         pci_write_config_word(dd->pcidev, pos +
1007                               ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
1008                               dd->ipath_msi_data);
1009         /* we restore the cachelinesize also, although it doesn't really
1010          * matter */
1011         pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE,
1012                               dd->ipath_pci_cacheline);
1013         /* and now set the pci master bit again */
1014         pci_set_master(dd->pcidev);
1015         ret = 1;
1016
1017 bail:
1018         return ret;
1019 }
1020
1021 /* This routine sleeps, so it can only be called from user context, not
1022  * from interrupt context.  If we need interrupt context, we can split
1023  * it into two routines.
1024 */
1025 static int ipath_setup_pe_reset(struct ipath_devdata *dd)
1026 {
1027         u64 val;
1028         int i;
1029         int ret;
1030
1031         /* Use ERROR so it shows up in logs, etc. */
1032         ipath_dev_err(dd, "Resetting InfiniPath unit %u\n", dd->ipath_unit);
1033         /* keep chip from being accessed in a few places */
1034         dd->ipath_flags &= ~(IPATH_INITTED|IPATH_PRESENT);
1035         val = dd->ipath_control | INFINIPATH_C_RESET;
1036         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, val);
1037         mb();
1038
1039         for (i = 1; i <= 5; i++) {
1040                 int r;
1041                 /* allow MBIST, etc. to complete; longer on each retry.
1042                  * We sometimes get machine checks from bus timeout if no
1043                  * response, so for now, make it *really* long.
1044                  */
1045                 msleep(1000 + (1 + i) * 2000);
1046                 if ((r =
1047                      pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
1048                                             dd->ipath_pcibar0)))
1049                         ipath_dev_err(dd, "rewrite of BAR0 failed: %d\n",
1050                                       r);
1051                 if ((r =
1052                      pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
1053                                             dd->ipath_pcibar1)))
1054                         ipath_dev_err(dd, "rewrite of BAR1 failed: %d\n",
1055                                       r);
1056                 /* now re-enable memory access */
1057                 if ((r = pci_enable_device(dd->pcidev)))
1058                         ipath_dev_err(dd, "pci_enable_device failed after "
1059                                       "reset: %d\n", r);
1060                 /* whether it worked or not, mark as present, again */
1061                 dd->ipath_flags |= IPATH_PRESENT;
1062                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
1063                 if (val == dd->ipath_revision) {
1064                         ipath_cdbg(VERBOSE, "Got matching revision "
1065                                    "register %llx on try %d\n",
1066                                    (unsigned long long) val, i);
1067                         ret = ipath_reinit_msi(dd);
1068                         goto bail;
1069                 }
1070                 /* Probably getting -1 back */
1071                 ipath_dbg("Didn't get expected revision register, "
1072                           "got %llx, try %d\n", (unsigned long long) val,
1073                           i + 1);
1074         }
1075         ret = 0; /* failed */
1076
1077 bail:
1078         return ret;
1079 }
1080
1081 /**
1082  * ipath_pe_put_tid - write a TID in chip
1083  * @dd: the infinipath device
1084  * @tidptr: pointer to the expected TID (in chip) to udpate
1085  * @tidtype: 0 for eager, 1 for expected
1086  * @pa: physical address of in memory buffer; ipath_tidinvalid if freeing
1087  *
1088  * This exists as a separate routine to allow for special locking etc.
1089  * It's used for both the full cleanup on exit, as well as the normal
1090  * setup and teardown.
1091  */
1092 static void ipath_pe_put_tid(struct ipath_devdata *dd, u64 __iomem *tidptr,
1093                              u32 type, unsigned long pa)
1094 {
1095         u32 __iomem *tidp32 = (u32 __iomem *)tidptr;
1096         unsigned long flags = 0; /* keep gcc quiet */
1097
1098         if (pa != dd->ipath_tidinvalid) {
1099                 if (pa & ((1U << 11) - 1)) {
1100                         dev_info(&dd->pcidev->dev, "BUG: physaddr %lx "
1101                                  "not 4KB aligned!\n", pa);
1102                         return;
1103                 }
1104                 pa >>= 11;
1105                 /* paranoia check */
1106                 if (pa & (7<<29))
1107                         ipath_dev_err(dd,
1108                                       "BUG: Physical page address 0x%lx "
1109                                       "has bits set in 31-29\n", pa);
1110
1111                 if (type == 0)
1112                         pa |= dd->ipath_tidtemplate;
1113                 else /* for now, always full 4KB page */
1114                         pa |= 2 << 29;
1115         }
1116
1117         /* workaround chip bug 9437 by writing each TID twice
1118          * and holding a spinlock around the writes, so they don't
1119          * intermix with other TID (eager or expected) writes
1120          * Unfortunately, this call can be done from interrupt level
1121          * for the port 0 eager TIDs, so we have to use irqsave
1122          */
1123         spin_lock_irqsave(&dd->ipath_tid_lock, flags);
1124         ipath_write_kreg(dd, dd->ipath_kregs->kr_scratch, 0xfeeddeaf);
1125         if (dd->ipath_kregbase)
1126                 writel(pa, tidp32);
1127         ipath_write_kreg(dd, dd->ipath_kregs->kr_scratch, 0xdeadbeef);
1128         mmiowb();
1129         spin_unlock_irqrestore(&dd->ipath_tid_lock, flags);
1130 }
1131 /**
1132  * ipath_pe_put_tid_2 - write a TID in chip, Revision 2 or higher
1133  * @dd: the infinipath device
1134  * @tidptr: pointer to the expected TID (in chip) to udpate
1135  * @tidtype: 0 for eager, 1 for expected
1136  * @pa: physical address of in memory buffer; ipath_tidinvalid if freeing
1137  *
1138  * This exists as a separate routine to allow for selection of the
1139  * appropriate "flavor". The static calls in cleanup just use the
1140  * revision-agnostic form, as they are not performance critical.
1141  */
1142 static void ipath_pe_put_tid_2(struct ipath_devdata *dd, u64 __iomem *tidptr,
1143                              u32 type, unsigned long pa)
1144 {
1145         u32 __iomem *tidp32 = (u32 __iomem *)tidptr;
1146
1147         if (pa != dd->ipath_tidinvalid) {
1148                 if (pa & ((1U << 11) - 1)) {
1149                         dev_info(&dd->pcidev->dev, "BUG: physaddr %lx "
1150                                  "not 2KB aligned!\n", pa);
1151                         return;
1152                 }
1153                 pa >>= 11;
1154                 /* paranoia check */
1155                 if (pa & (7<<29))
1156                         ipath_dev_err(dd,
1157                                       "BUG: Physical page address 0x%lx "
1158                                       "has bits set in 31-29\n", pa);
1159
1160                 if (type == 0)
1161                         pa |= dd->ipath_tidtemplate;
1162                 else /* for now, always full 4KB page */
1163                         pa |= 2 << 29;
1164         }
1165         if (dd->ipath_kregbase)
1166                 writel(pa, tidp32);
1167         mmiowb();
1168 }
1169
1170
1171 /**
1172  * ipath_pe_clear_tid - clear all TID entries for a port, expected and eager
1173  * @dd: the infinipath device
1174  * @port: the port
1175  *
1176  * clear all TID entries for a port, expected and eager.
1177  * Used from ipath_close().  On this chip, TIDs are only 32 bits,
1178  * not 64, but they are still on 64 bit boundaries, so tidbase
1179  * is declared as u64 * for the pointer math, even though we write 32 bits
1180  */
1181 static void ipath_pe_clear_tids(struct ipath_devdata *dd, unsigned port)
1182 {
1183         u64 __iomem *tidbase;
1184         unsigned long tidinv;
1185         int i;
1186
1187         if (!dd->ipath_kregbase)
1188                 return;
1189
1190         ipath_cdbg(VERBOSE, "Invalidate TIDs for port %u\n", port);
1191
1192         tidinv = dd->ipath_tidinvalid;
1193         tidbase = (u64 __iomem *)
1194                 ((char __iomem *)(dd->ipath_kregbase) +
1195                  dd->ipath_rcvtidbase +
1196                  port * dd->ipath_rcvtidcnt * sizeof(*tidbase));
1197
1198         for (i = 0; i < dd->ipath_rcvtidcnt; i++)
1199                 ipath_pe_put_tid(dd, &tidbase[i], 0, tidinv);
1200
1201         tidbase = (u64 __iomem *)
1202                 ((char __iomem *)(dd->ipath_kregbase) +
1203                  dd->ipath_rcvegrbase +
1204                  port * dd->ipath_rcvegrcnt * sizeof(*tidbase));
1205
1206         for (i = 0; i < dd->ipath_rcvegrcnt; i++)
1207                 ipath_pe_put_tid(dd, &tidbase[i], 1, tidinv);
1208 }
1209
1210 /**
1211  * ipath_pe_tidtemplate - setup constants for TID updates
1212  * @dd: the infinipath device
1213  *
1214  * We setup stuff that we use a lot, to avoid calculating each time
1215  */
1216 static void ipath_pe_tidtemplate(struct ipath_devdata *dd)
1217 {
1218         u32 egrsize = dd->ipath_rcvegrbufsize;
1219
1220         /* For now, we always allocate 4KB buffers (at init) so we can
1221          * receive max size packets.  We may want a module parameter to
1222          * specify 2KB or 4KB and/or make be per port instead of per device
1223          * for those who want to reduce memory footprint.  Note that the
1224          * ipath_rcvhdrentsize size must be large enough to hold the largest
1225          * IB header (currently 96 bytes) that we expect to handle (plus of
1226          * course the 2 dwords of RHF).
1227          */
1228         if (egrsize == 2048)
1229                 dd->ipath_tidtemplate = 1U << 29;
1230         else if (egrsize == 4096)
1231                 dd->ipath_tidtemplate = 2U << 29;
1232         else {
1233                 egrsize = 4096;
1234                 dev_info(&dd->pcidev->dev, "BUG: unsupported egrbufsize "
1235                          "%u, using %u\n", dd->ipath_rcvegrbufsize,
1236                          egrsize);
1237                 dd->ipath_tidtemplate = 2U << 29;
1238         }
1239         dd->ipath_tidinvalid = 0;
1240 }
1241
1242 static int ipath_pe_early_init(struct ipath_devdata *dd)
1243 {
1244         dd->ipath_flags |= IPATH_4BYTE_TID;
1245
1246         /*
1247          * For openfabrics, we need to be able to handle an IB header of
1248          * 24 dwords.  HT chip has arbitrary sized receive buffers, so we
1249          * made them the same size as the PIO buffers.  This chip does not
1250          * handle arbitrary size buffers, so we need the header large enough
1251          * to handle largest IB header, but still have room for a 2KB MTU
1252          * standard IB packet.
1253          */
1254         dd->ipath_rcvhdrentsize = 24;
1255         dd->ipath_rcvhdrsize = IPATH_DFLT_RCVHDRSIZE;
1256
1257         /*
1258          * To truly support a 4KB MTU (for usermode), we need to
1259          * bump this to a larger value.  For now, we use them for
1260          * the kernel only.
1261          */
1262         dd->ipath_rcvegrbufsize = 2048;
1263         /*
1264          * the min() check here is currently a nop, but it may not always
1265          * be, depending on just how we do ipath_rcvegrbufsize
1266          */
1267         dd->ipath_ibmaxlen = min(dd->ipath_piosize2k,
1268                                  dd->ipath_rcvegrbufsize +
1269                                  (dd->ipath_rcvhdrentsize << 2));
1270         dd->ipath_init_ibmaxlen = dd->ipath_ibmaxlen;
1271
1272         /*
1273          * We can request a receive interrupt for 1 or
1274          * more packets from current offset.  For now, we set this
1275          * up for a single packet.
1276          */
1277         dd->ipath_rhdrhead_intr_off = 1ULL<<32;
1278
1279         ipath_get_eeprom_info(dd);
1280
1281         return 0;
1282 }
1283
1284 int __attribute__((weak)) ipath_unordered_wc(void)
1285 {
1286         return 0;
1287 }
1288
1289 /**
1290  * ipath_init_pe_get_base_info - set chip-specific flags for user code
1291  * @pd: the infinipath port
1292  * @kbase: ipath_base_info pointer
1293  *
1294  * We set the PCIE flag because the lower bandwidth on PCIe vs
1295  * HyperTransport can affect some user packet algorithims.
1296  */
1297 static int ipath_pe_get_base_info(struct ipath_portdata *pd, void *kbase)
1298 {
1299         struct ipath_base_info *kinfo = kbase;
1300         struct ipath_devdata *dd;
1301
1302         if (ipath_unordered_wc()) {
1303                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_FORCE_WC_ORDER;
1304                 ipath_cdbg(PROC, "Intel processor, forcing WC order\n");
1305         }
1306         else
1307                 ipath_cdbg(PROC, "Not Intel processor, WC ordered\n");
1308
1309         if (pd == NULL)
1310                 goto done;
1311
1312         dd = pd->port_dd;
1313
1314         if (dd != NULL && dd->ipath_minrev >= 2) {
1315                 ipath_cdbg(PROC, "IBA6120 Rev2, allow multiple PBC write\n");
1316                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_PBC_REWRITE;
1317                 ipath_cdbg(PROC, "IBA6120 Rev2, allow loose DMA alignment\n");
1318                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_LOOSE_DMA_ALIGN;
1319         }
1320
1321 done:
1322         kinfo->spi_runtime_flags |= IPATH_RUNTIME_PCIE;
1323         return 0;
1324 }
1325
1326 /**
1327  * ipath_init_iba6120_funcs - set up the chip-specific function pointers
1328  * @dd: the infinipath device
1329  *
1330  * This is global, and is called directly at init to set up the
1331  * chip-specific function pointers for later use.
1332  */
1333 void ipath_init_iba6120_funcs(struct ipath_devdata *dd)
1334 {
1335         dd->ipath_f_intrsetup = ipath_pe_intconfig;
1336         dd->ipath_f_bus = ipath_setup_pe_config;
1337         dd->ipath_f_reset = ipath_setup_pe_reset;
1338         dd->ipath_f_get_boardname = ipath_pe_boardname;
1339         dd->ipath_f_init_hwerrors = ipath_pe_init_hwerrors;
1340         dd->ipath_f_early_init = ipath_pe_early_init;
1341         dd->ipath_f_handle_hwerrors = ipath_pe_handle_hwerrors;
1342         dd->ipath_f_quiet_serdes = ipath_pe_quiet_serdes;
1343         dd->ipath_f_bringup_serdes = ipath_pe_bringup_serdes;
1344         dd->ipath_f_clear_tids = ipath_pe_clear_tids;
1345         if (dd->ipath_minrev >= 2)
1346                 dd->ipath_f_put_tid = ipath_pe_put_tid_2;
1347         else
1348                 dd->ipath_f_put_tid = ipath_pe_put_tid;
1349         dd->ipath_f_cleanup = ipath_setup_pe_cleanup;
1350         dd->ipath_f_setextled = ipath_setup_pe_setextled;
1351         dd->ipath_f_get_base_info = ipath_pe_get_base_info;
1352
1353         /* initialize chip-specific variables */
1354         dd->ipath_f_tidtemplate = ipath_pe_tidtemplate;
1355
1356         /*
1357          * setup the register offsets, since they are different for each
1358          * chip
1359          */
1360         dd->ipath_kregs = &ipath_pe_kregs;
1361         dd->ipath_cregs = &ipath_pe_cregs;
1362
1363         ipath_init_pe_variables(dd);
1364 }
1365